---

ROAD TRAFFIC ACCIDENT DATASET OF ADDIS ABABA CITY

---

🧾Description: This data set is collected from Addis Ababa Sub-city police departments for master's research work. The data set has been prepared from manual records of road traffic accidents of the year 2017-20. All the sensitive information has been excluded during data encoding and finally it has 32 features and 12316 instances of the accident. Then it is preprocessed and for identification of major causes of the accident by analyzing it using different machine learning classification algorithms.

source of dataset - Click Here

🧭 Problem Statement: The target feature is Accident_severity which is a multi-class variable. The task is to classify this variable based on the other 31 features step-by-step by going through each day's task. Your metric for evaluation will be f1-score

---

Libraries

---

import pandas as pd
import numpy as np
from ydata_profiling import ProfileReport
import matplotlib.pyplot as plt
import plotly.graph_objects as go
from plotly.subplots import make_subplots
import plotly.express as px
import plotly.offline as py
import plotly.tools as tls
import seaborn as sns
import matplotlib
from collections import Counter
import joblib
import shap

from dataprep.eda import plot, plot_correlation, create_report, plot_missing
from IPython.display import display_html
from EDA import DataExplorer as de

pd.set_option('display.max_columns', None)

import warnings
warnings.filterwarnings('ignore')
matplotlib.style.use('seaborn')

---

Data

---

df = pd.read_csv("data/RTA dataset.csv")

display_html(df.head(2), df.tail(2), df.sample(2))

	Time	Day_of_week	Age_band_of_driver	Sex_of_driver	Educational_level	Vehicle_driver_relation	Driving_experience	Type_of_vehicle	Owner_of_vehicle	Service_year_of_vehicle	Defect_of_vehicle	Area_accident_occured	Lanes_or_Medians	Road_allignment	Types_of_Junction	Road_surface_type	Road_surface_conditions	Light_conditions	Weather_conditions	Type_of_collision	Number_of_vehicles_involved	Number_of_casualties	Vehicle_movement	Casualty_class	Sex_of_casualty	Age_band_of_casualty	Casualty_severity	Work_of_casuality	Fitness_of_casuality	Pedestrian_movement	Cause_of_accident	Accident_severity
0	17:02:00	Monday	18-30	Male	Above high school	Employee	1-2yr	Automobile	Owner	Above 10yr	No defect	Residential areas	NaN	Tangent road with flat terrain	No junction	Asphalt roads	Dry	Daylight	Normal	Collision with roadside-parked vehicles	2	2	Going straight	na	na	na	na	NaN	NaN	Not a Pedestrian	Moving Backward	Slight Injury
1	17:02:00	Monday	31-50	Male	Junior high school	Employee	Above 10yr	Public (> 45 seats)	Owner	5-10yrs	No defect	Office areas	Undivided Two way	Tangent road with flat terrain	No junction	Asphalt roads	Dry	Daylight	Normal	Vehicle with vehicle collision	2	2	Going straight	na	na	na	na	NaN	NaN	Not a Pedestrian	Overtaking	Slight Injury

	Time	Day_of_week	Age_band_of_driver	Sex_of_driver	Educational_level	Vehicle_driver_relation	Driving_experience	Type_of_vehicle	Owner_of_vehicle	Service_year_of_vehicle	Defect_of_vehicle	Area_accident_occured	Lanes_or_Medians	Road_allignment	Types_of_Junction	Road_surface_type	Road_surface_conditions	Light_conditions	Weather_conditions	Type_of_collision	Number_of_vehicles_involved	Number_of_casualties	Vehicle_movement	Casualty_class	Sex_of_casualty	Age_band_of_casualty	Casualty_severity	Work_of_casuality	Fitness_of_casuality	Pedestrian_movement	Cause_of_accident	Accident_severity
12314	13:55:00	Sunday	18-30	Female	Junior high school	Employee	Above 10yr	Lorry (41?100Q)	Owner	2-5yrs	No defect	Office areas	Undivided Two way	Tangent road with mountainous terrain and	No junction	Asphalt roads	Dry	Darkness - lights lit	Normal	Vehicle with vehicle collision	2	1	Other	na	na	na	na	Driver	Normal	Not a Pedestrian	Driving under the influence of drugs	Slight Injury
12315	13:55:00	Sunday	18-30	Male	Junior high school	Employee	5-10yr	Other	Owner	2-5yrs	No defect	Outside rural areas	Undivided Two way	Tangent road with mountainous terrain and	O Shape	Asphalt roads	Dry	Darkness - lights lit	Normal	Vehicle with vehicle collision	2	1	Stopping	Pedestrian	Female	5	3	Driver	Normal	Crossing from nearside - masked by parked or s...	Changing lane to the right	Slight Injury

	Time	Day_of_week	Age_band_of_driver	Sex_of_driver	Educational_level	Vehicle_driver_relation	Driving_experience	Type_of_vehicle	Owner_of_vehicle	Service_year_of_vehicle	Defect_of_vehicle	Area_accident_occured	Lanes_or_Medians	Road_allignment	Types_of_Junction	Road_surface_type	Road_surface_conditions	Light_conditions	Weather_conditions	Type_of_collision	Number_of_vehicles_involved	Number_of_casualties	Vehicle_movement	Casualty_class	Sex_of_casualty	Age_band_of_casualty	Casualty_severity	Work_of_casuality	Fitness_of_casuality	Pedestrian_movement	Cause_of_accident	Accident_severity
1045	13:55:00	Sunday	31-50	Male	NaN	NaN	NaN	NaN	NaN	NaN	NaN	Unknown	other	Tangent road with flat terrain	Y Shape	Asphalt roads	Dry	Daylight	Normal	Vehicle with vehicle collision	2	4	Going straight	na	na	na	na	NaN	NaN	Not a Pedestrian	No priority to vehicle	Slight Injury
9368	8:00:00	Monday	18-30	Male	Junior high school	Employee	Below 1yr	Lorry (41?100Q)	Owner	2-5yrs	No defect	Residential areas	Double carriageway (median)	Tangent road with flat terrain	Y Shape	Asphalt roads	Dry	Daylight	Normal	Rollover	1	1	Going straight	na	na	na	na	NaN	NaN	Not a Pedestrian	No priority to vehicle	Slight Injury

# ProfileReport(df)
# create_report(df)

df.shape

(12316, 32)

df.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 12316 entries, 0 to 12315
Data columns (total 32 columns):
 #   Column                       Non-Null Count  Dtype 
---  ------                       --------------  ----- 
 0   Time                         12316 non-null  object
 1   Day_of_week                  12316 non-null  object
 2   Age_band_of_driver           12316 non-null  object
 3   Sex_of_driver                12316 non-null  object
 4   Educational_level            11575 non-null  object
 5   Vehicle_driver_relation      11737 non-null  object
 6   Driving_experience           11487 non-null  object
 7   Type_of_vehicle              11366 non-null  object
 8   Owner_of_vehicle             11834 non-null  object
 9   Service_year_of_vehicle      8388 non-null   object
 10  Defect_of_vehicle            7889 non-null   object
 11  Area_accident_occured        12077 non-null  object
 12  Lanes_or_Medians             11931 non-null  object
 13  Road_allignment              12174 non-null  object
 14  Types_of_Junction            11429 non-null  object
 15  Road_surface_type            12144 non-null  object
 16  Road_surface_conditions      12316 non-null  object
 17  Light_conditions             12316 non-null  object
 18  Weather_conditions           12316 non-null  object
 19  Type_of_collision            12161 non-null  object
 20  Number_of_vehicles_involved  12316 non-null  int64 
 21  Number_of_casualties         12316 non-null  int64 
 22  Vehicle_movement             12008 non-null  object
 23  Casualty_class               12316 non-null  object
 24  Sex_of_casualty              12316 non-null  object
 25  Age_band_of_casualty         12316 non-null  object
 26  Casualty_severity            12316 non-null  object
 27  Work_of_casuality            9118 non-null   object
 28  Fitness_of_casuality         9681 non-null   object
 29  Pedestrian_movement          12316 non-null  object
 30  Cause_of_accident            12316 non-null  object
 31  Accident_severity            12316 non-null  object
dtypes: int64(2), object(30)
memory usage: 3.0+ MB

df.columns

Index(['Time', 'Day_of_week', 'Age_band_of_driver', 'Sex_of_driver',
       'Educational_level', 'Vehicle_driver_relation', 'Driving_experience',
       'Type_of_vehicle', 'Owner_of_vehicle', 'Service_year_of_vehicle',
       'Defect_of_vehicle', 'Area_accident_occured', 'Lanes_or_Medians',
       'Road_allignment', 'Types_of_Junction', 'Road_surface_type',
       'Road_surface_conditions', 'Light_conditions', 'Weather_conditions',
       'Type_of_collision', 'Number_of_vehicles_involved',
       'Number_of_casualties', 'Vehicle_movement', 'Casualty_class',
       'Sex_of_casualty', 'Age_band_of_casualty', 'Casualty_severity',
       'Work_of_casuality', 'Fitness_of_casuality', 'Pedestrian_movement',
       'Cause_of_accident', 'Accident_severity'],
      dtype='object')

col_map={
    'Time': 'time',
    'Day_of_week': 'day_of_week',
    'Age_band_of_driver': 'driver_age',
    'Sex_of_driver': 'driver_sex',
    'Educational_level': 'educational_level',
    'Vehicle_driver_relation': 'vehicle_driver_relation',
    'Driving_experience': 'driving_experience',
    'Type_of_vehicle': 'vehicle_type',
    'Owner_of_vehicle': 'vehicle_owner',
    'Service_year_of_vehicle': 'service_year',
    'Defect_of_vehicle': 'vehicle_defect',
    'Area_accident_occured': 'accident_area',
    'Lanes_or_Medians': 'lanes',
    'Road_allignment': 'road_allignment',
    'Types_of_Junction': 'junction_type',
    'Road_surface_type': 'surface_type',
    'Road_surface_conditions': 'road_surface_conditions',
    'Light_conditions': 'light_condition',
    'Weather_conditions': 'weather_condition',
    'Type_of_collision': 'collision_type',
    'Number_of_vehicles_involved': 'vehicles_involved',
    'Number_of_casualties': 'casualties',
    'Vehicle_movement': 'vehicle_movement',
    'Casualty_class': 'casualty_class',
    'Sex_of_casualty': 'casualty_sex' , 
    'Age_band_of_casualty': 'casualty_age', 
    'Casualty_severity': 'casualty_severity',
    'Work_of_casuality': 'casualty_work', 
    'Fitness_of_casuality': 'casualty_fitness',
    'Pedestrian_movement': 'pedestrian_movement',
    'Cause_of_accident': 'accident_cause',
    'Accident_severity': 'accident_severity'
}
df.rename(columns=col_map, inplace=True)

# calculate the missing values count and convert it to a dataframe
missing_vals = pd.DataFrame(df.isna().sum(), columns=['missing'])

# concatenate the two dataframes
result = pd.concat([df.describe(include=['O']).T, missing_vals], axis=1)

result

	count	unique	top	freq	missing
time	12316	1074	15:30:00	120	0
day_of_week	12316	7	Friday	2041	0
driver_age	12316	5	18-30	4271	0
driver_sex	12316	3	Male	11437	0
educational_level	11575	7	Junior high school	7619	741
vehicle_driver_relation	11737	4	Employee	9627	579
driving_experience	11487	7	5-10yr	3363	829
vehicle_type	11366	17	Automobile	3205	950
vehicle_owner	11834	4	Owner	10459	482
service_year	8388	6	Unknown	2883	3928
vehicle_defect	7889	3	No defect	7777	4427
accident_area	12077	14	Other	3819	239
lanes	11931	7	Two-way (divided with broken lines road marking)	4411	385
road_allignment	12174	9	Tangent road with flat terrain	10459	142
junction_type	11429	8	Y Shape	4543	887
surface_type	12144	5	Asphalt roads	11296	172
road_surface_conditions	12316	4	Dry	9340	0
light_condition	12316	4	Daylight	8798	0
weather_condition	12316	9	Normal	10063	0
collision_type	12161	10	Vehicle with vehicle collision	8774	155
vehicle_movement	12008	13	Going straight	8158	308
casualty_class	12316	4	Driver or rider	4944	0
casualty_sex	12316	3	Male	5253	0
casualty_age	12316	6	na	4443	0
casualty_severity	12316	4	3	7076	0
casualty_work	9118	7	Driver	5903	3198
casualty_fitness	9681	5	Normal	9608	2635
pedestrian_movement	12316	9	Not a Pedestrian	11390	0
accident_cause	12316	20	No distancing	2263	0
accident_severity	12316	3	Slight Injury	10415	0
vehicles_involved	NaN	NaN	NaN	NaN	0
casualties	NaN	NaN	NaN	NaN	0

de.generate_data_description_table(df)

	count	mean	std	min	25%	50%	75%	max
casualties	12316.000000	1.548149	1.007179	1.000000	1.000000	1.000000	2.000000	8.000000
vehicles_involved	12316.000000	2.040679	0.688790	1.000000	2.000000	2.000000	2.000000	7.000000

print("Number of duplicates: ", df.duplicated().sum())

Number of duplicates:  0

for i in df.columns:
    print(f"Unique value in {i}:")
    print(df[i].unique(),'\n')

Unique value in time:
['17:02:00' '1:06:00' '14:15:00' ... '7:24:00' '19:18:00' '2:47:00'] 

Unique value in day_of_week:
['Monday' 'Sunday' 'Friday' 'Wednesday' 'Saturday' 'Thursday' 'Tuesday'] 

Unique value in driver_age:
['18-30' '31-50' 'Under 18' 'Over 51' 'Unknown'] 

Unique value in driver_sex:
['Male' 'Female' 'Unknown'] 

Unique value in educational_level:
['Above high school' 'Junior high school' nan 'Elementary school'
 'High school' 'Unknown' 'Illiterate' 'Writing & reading'] 

Unique value in vehicle_driver_relation:
['Employee' 'Unknown' 'Owner' nan 'Other'] 

Unique value in driving_experience:
['1-2yr' 'Above 10yr' '5-10yr' '2-5yr' nan 'No Licence' 'Below 1yr'
 'unknown'] 

Unique value in vehicle_type:
['Automobile' 'Public (> 45 seats)' 'Lorry (41?100Q)' nan
 'Public (13?45 seats)' 'Lorry (11?40Q)' 'Long lorry' 'Public (12 seats)'
 'Taxi' 'Pick up upto 10Q' 'Stationwagen' 'Ridden horse' 'Other' 'Bajaj'
 'Turbo' 'Motorcycle' 'Special vehicle' 'Bicycle'] 

Unique value in vehicle_owner:
['Owner' 'Governmental' nan 'Organization' 'Other'] 

Unique value in service_year:
['Above 10yr' '5-10yrs' nan '1-2yr' '2-5yrs' 'Unknown' 'Below 1yr'] 

Unique value in vehicle_defect:
['No defect' nan '7' '5'] 

Unique value in accident_area:
['Residential areas' 'Office areas' '  Recreational areas'
 ' Industrial areas' nan 'Other' ' Church areas' '  Market areas'
 'Unknown' 'Rural village areas' ' Outside rural areas' ' Hospital areas'
 'School areas' 'Rural village areasOffice areas' 'Recreational areas'] 

Unique value in lanes:
[nan 'Undivided Two way' 'other' 'Double carriageway (median)' 'One way'
 'Two-way (divided with solid lines road marking)'
 'Two-way (divided with broken lines road marking)' 'Unknown'] 

Unique value in road_allignment:
['Tangent road with flat terrain' nan
 'Tangent road with mild grade and flat terrain' 'Escarpments'
 'Tangent road with rolling terrain' 'Gentle horizontal curve'
 'Tangent road with mountainous terrain and'
 'Steep grade downward with mountainous terrain' 'Sharp reverse curve'
 'Steep grade upward with mountainous terrain'] 

Unique value in junction_type:
['No junction' 'Y Shape' 'Crossing' 'O Shape' 'Other' 'Unknown' 'T Shape'
 'X Shape' nan] 

Unique value in surface_type:
['Asphalt roads' 'Earth roads' nan 'Asphalt roads with some distress'
 'Gravel roads' 'Other'] 

Unique value in road_surface_conditions:
['Dry' 'Wet or damp' 'Snow' 'Flood over 3cm. deep'] 

Unique value in light_condition:
['Daylight' 'Darkness - lights lit' 'Darkness - no lighting'
 'Darkness - lights unlit'] 

Unique value in weather_condition:
['Normal' 'Raining' 'Raining and Windy' 'Cloudy' 'Other' 'Windy' 'Snow'
 'Unknown' 'Fog or mist'] 

Unique value in collision_type:
['Collision with roadside-parked vehicles'
 'Vehicle with vehicle collision' 'Collision with roadside objects'
 'Collision with animals' 'Other' 'Rollover' 'Fall from vehicles'
 'Collision with pedestrians' 'With Train' 'Unknown' nan] 

Unique value in vehicles_involved:
[2 1 3 6 4 7] 

Unique value in casualties:
[2 1 3 4 6 5 8 7] 

Unique value in vehicle_movement:
['Going straight' 'U-Turn' 'Moving Backward' 'Turnover' 'Waiting to go'
 'Getting off' 'Reversing' 'Unknown' 'Parked' 'Stopping' 'Overtaking'
 'Other' 'Entering a junction' nan] 

Unique value in casualty_class:
['na' 'Driver or rider' 'Pedestrian' 'Passenger'] 

Unique value in casualty_sex:
['na' 'Male' 'Female'] 

Unique value in casualty_age:
['na' '31-50' '18-30' 'Under 18' 'Over 51' '5'] 

Unique value in casualty_severity:
['na' '3' '2' '1'] 

Unique value in casualty_work:
[nan 'Driver' 'Other' 'Unemployed' 'Employee' 'Self-employed' 'Student'
 'Unknown'] 

Unique value in casualty_fitness:
[nan 'Normal' 'Deaf' 'Other' 'Blind' 'NormalNormal'] 

Unique value in pedestrian_movement:
['Not a Pedestrian' "Crossing from driver's nearside"
 'Crossing from nearside - masked by parked or statioNot a Pedestrianry vehicle'
 'Unknown or other'
 'Crossing from offside - masked by  parked or statioNot a Pedestrianry vehicle'
 'In carriageway, statioNot a Pedestrianry - not crossing  (standing or playing)'
 'Walking along in carriageway, back to traffic'
 'Walking along in carriageway, facing traffic'
 'In carriageway, statioNot a Pedestrianry - not crossing  (standing or playing) - masked by parked or statioNot a Pedestrianry vehicle'] 

Unique value in accident_cause:
['Moving Backward' 'Overtaking' 'Changing lane to the left'
 'Changing lane to the right' 'Overloading' 'Other'
 'No priority to vehicle' 'No priority to pedestrian' 'No distancing'
 'Getting off the vehicle improperly' 'Improper parking' 'Overspeed'
 'Driving carelessly' 'Driving at high speed' 'Driving to the left'
 'Unknown' 'Overturning' 'Turnover' 'Driving under the influence of drugs'
 'Drunk driving'] 

Unique value in accident_severity:
['Slight Injury' 'Serious Injury' 'Fatal injury'] 

# df = df.replace('na', np.nan)

class_count = pd.value_counts(df['accident_severity'], sort = True).sort_index()

fig = go.Figure(data=[go.Pie(labels=['Fatal injury', 'Serious Injury', 'Slight Injury'], 
                                    values=class_count, 
                                    pull=[0.1, 0, 0], 
                                    opacity=0.85)])
fig.update_layout(
    title_text="Composition of Accident Severity")
fig.show()

fig = px.histogram(df, x="accident_severity",
                   title='Composition of Accident Severity',
                   opacity=0.85, # represent bars with log scale
                   color = 'accident_severity', text_auto=True)

fig.update_layout({
"plot_bgcolor": "rgba(0, 0, 0, 0)",
"paper_bgcolor": "rgba(0, 0, 0, 0)",
})

fig.show()

fig = px.treemap(df, path=['accident_cause'], width=800, height=400)
fig.update_layout(
    margin = dict(t=0, l=0, r=0, b=50))

fig.show()

# converting 'time' to datetime
df['time'] = pd.to_datetime(df['time'])

# date (day-month-year) time
df["time"].dt.hour

# extracting hour and minute from timestamp
df['hour'] = df['time'].dt.hour
df['minute'] = df['time'].dt.minute
df.drop('time', axis=1, inplace=True)

plt.figure(figsize=(15,70))
plotnumber = 1

for col in df.drop(['hour', 'minute', 'lanes', 'road_allignment', 'pedestrian_movement'], axis=1):
    if plotnumber <= df.shape[1]:
        ax1 = plt.subplot(16,2,plotnumber)
        sns.countplot(data=df, y=col, palette='Dark2')
        plt.xticks(fontsize=12)
        plt.yticks(fontsize=12)
        plt.title(col.title(), fontsize=14)
        plt.xlabel('')
        plt.ylabel('')
    plotnumber +=1
plt.tight_layout()

plt.figure(figsize=(10,3))
sns.countplot(data=df, y='lanes', palette = 'Dark2')
plt.xticks(fontsize=12)
plt.yticks(fontsize=12)
plt.title('Lanes', fontsize=14)
plt.xlabel('')
plt.ylabel('')
plt.tight_layout()

plt.figure(figsize=(10,3))
sns.countplot(data=df, y='road_allignment', palette = 'Dark2')
plt.xticks(fontsize=12)
plt.yticks(fontsize=12)
plt.title('Road Allignment', fontsize=14)
plt.xlabel('')
plt.ylabel('')
plt.tight_layout()

plt.figure(figsize=(10,5))
sns.countplot(data=df, y='pedestrian_movement', palette = 'Dark2')
plt.xticks(fontsize=12)
plt.yticks(fontsize=12)
plt.title('Pedestrian Movement', fontsize=14)
plt.xlabel('')
plt.ylabel('')
plt.tight_layout()

plt.figure(figsize=(10,5))
sns.countplot(data=df, y='hour')
plt.xticks(fontsize=12)
plt.yticks(fontsize=12)
plt.title('Hour', fontsize=14)
plt.xlabel('')
plt.ylabel('')
plt.tight_layout()

plt.figure(figsize=(10,15))
sns.countplot(data=df, y='minute')
plt.xticks(fontsize=12)
plt.yticks(fontsize=12)
plt.title('Minute', fontsize=14)
plt.xlabel('')
plt.ylabel('')
plt.tight_layout()

💡 Observations:

Most of the accidents:

• involved 2 vehicles and 2 casualties
• occured on Fridays and after noon hours

Most of the drivers:

• are male and in 18-30 yrs age group
• have only went upto Junior high school and are employees
• have 5-10 yrs of driving experience

Most of the accidents happened with personally owned passenger vehicles.

Most of the drivers have met with accident on:

• two-way lanes
• tangent road with flat terrains
• Y shaped junctions
• asphalt roads
• day time
• normal weather conditions

Most of the casualties:

• happened to physically fit male drivers
• are of severity 3

The conditions on which most of the drivers met with the accident are:

• vehicle to vehicle collision
• straight movement
• no pedestrian accidents

Not keeping enough distance between the vehicles was the major cause for most of the accidents and majority of the accidents resulted in slight injury.

💡 Observations:

• Most of the accidents have taken place between 3pm to 6pm with maximum numbers at 5 pm
• Most of the time in minutes seems to have been recorded in intervals of 5 minutes for the sake of convenience. Hence more numbers to be seen at the 0 and 30 minutes.

min = list(range(5,56, 5))
def convert_minutes(x: int):
    for m in min:
        if x % m == x and x > m-5:
            return m
        if x in [56,57,58,59]:
            return 0
        if x in min+[0]:
            return x

df['minute'] = df['minute'].apply(lambda x: convert_minutes(x))

plt.figure(figsize=(5,7))
sns.countplot(data=df, y='minute')
plt.xticks(fontsize=12)
plt.yticks(fontsize=12)
plt.title('Minute', fontsize=14)
plt.xlabel('')
plt.ylabel('')
plt.tight_layout()

💡 Observations:

• Most of the accidents with fatal injuries happened between 2pm to 7pm.
• Most of the accidents with fatal injuries happened on weekends.
• Highest number of non-fatal injuries happened at 5pm.
• Highest number of non-fatal injuries happened on fridays.
• Most accidents are caused by drivers aged 18-30 and the least by drivers aged under 18.
• Proportion of fatal accidents are lower for female drivers.
• Drivers with 2-5yrs of experience caused most accidents with fatal injury and those with 5-10yrs experience caused most accidents with non-fatal injuries.
• Proportion of fatal injuries caused by vehicles with more than 10yrs of service is lower compared to non-fatal injuries.
• Though most of the accident happened around offices, a higher proportion of accidents happened around residential areas have led to fatal injuries.
• Most accidents with fatal injuries occured on undivided two-ways.
• Most accidents with non-fatal injuries occured on two-way divided with broken lines road marking
• Double carriageway has a lower proportion of fatal accidents compared to non-fatal accidents.
• Severity of accidents increases at places with no junctions and the most number of fatal injuries occured at places with no junctions.
• Crossings and Y-shaped junctions shows a decreasing trend with respect to severity of accidents though the numbers are high.
• Most of the accidents with fatal injury have happened at night.
• Collisions with pedestrians have resulted in more fatal injuries compared to non-fatal injuries.
• Accidents with 4 casualties have a huge proportion on fatal injuries compared to non-fatal injuries.
• Accidents involving drivers with 18-30yrs of age have an increasing pattern with respect to severity of injuries.
• Moving backward led to most of the accidents with fatal injuries whereas failing to keep enough distance between vehicles led to more number of accidents with non-fatal injuries.

Testing Hypothesis

hypothesis_df = df.copy(deep=True)

❗️The copy() method performs a shallow copy, which means that the new object references the same data as the original object. This can cause issues if the original object is modified, as the changes will also affect the new object.

To avoid this issue, the deep=True parameter is used to perform a deep copy, which means that a completely new object is created with its own data. This ensures that any changes made to the original object do not affect the new object.

Hypothesis 1: Which of both genders is actually dangerous in Driving?

Calculating the Conditional Probability; Probability of an Accident Happen When the Driver is a Female and Repeating the Calculations when it's Male.
P(Gender = Male | Severity = 2) P(Gender = Female | Severity = 2)

Counter(df["driver_sex"])

Counter({'Male': 11437, 'Female': 701, 'Unknown': 178})

((hypothesis_df.groupby(['driver_sex']).size() / hypothesis_df["driver_sex"].count()) * 100).add_prefix('Accidents(in %) Caused by ')

driver_sex
Accidents(in %) Caused by Female      5.691783
Accidents(in %) Caused by Male       92.862943
Accidents(in %) Caused by Unknown     1.445274
dtype: float64

male_df = hypothesis_df.loc[hypothesis_df.driver_sex == 'Male']
female_df = hypothesis_df.loc[hypothesis_df.driver_sex == 'Female']

fig = go.Figure()

fig.add_trace(go.Bar(x=male_df.accident_cause.unique(), y=male_df.accident_cause.value_counts().values, 
                     name='Male'))
fig.add_trace(go.Bar(x=female_df.accident_cause.unique(), y=female_df.accident_cause.value_counts().values, 
                     name='Female'))

fig.update_layout(title='Accident Cause by Driver Sex',
                  xaxis_title='Accident Cause', yaxis_title='Count',
                  barmode='group', bargap=0.1, bargroupgap=0.2)

fig.show()

injury_by_sex = pd.crosstab(index=hypothesis_df['driver_sex'].loc[hypothesis_df['driver_sex'] !="Unknown"], columns=df['accident_severity']
                 , margins=True)
injury_by_sex

accident_severity	Fatal injury	Serious Injury	Slight Injury	All
driver_sex
Female	5	104	592	701
Male	152	1621	9664	11437
All	157	1725	10256	12138

$P(Fatal | Male) = P(Fatal | Female)?$

print("P(Fatal | Female) = {:.2f}%".format(injury_by_sex.iloc[0,0] / injury_by_sex.iloc[0, 3]*100))
print("P(Fatal | Male) = {:.2f}%\n".format(injury_by_sex.iloc[1,0] / injury_by_sex.iloc[1, 3]*100))

P(Fatal | Female) = 0.71%
P(Fatal | Male) = 1.33%

$\therefore P(Fatal | Male) \approx 2P(Fatal | Female)$

print("P(Serious | Female) = {:.2f}%".format(injury_by_sex.iloc[0,1] / injury_by_sex.iloc[0, 3]*100))
print("P(Serious | Male) = {:.2f}%\n".format(injury_by_sex.iloc[1,1] / injury_by_sex.iloc[1, 3]*100))

P(Serious | Female) = 14.84%
P(Serious | Male) = 14.17%

print("P(Slight | Female) = {:.2f}%".format(injury_by_sex.iloc[0,2] / injury_by_sex.iloc[0, 3]*100))
print("P(Slight | Male) = {:.2f}%\n".format(injury_by_sex.iloc[1,2] / injury_by_sex.iloc[1, 3]*100))

P(Slight | Female) = 84.45%
P(Slight | Male) = 84.50%

print("P(Gender = Female | Severity = Fatal) = {:.2f}%".format(injury_by_sex.iloc[0,0]/injury_by_sex.iloc[2,0]*100))
print("P(Gender = Male   | Severity = Fatal) = {:.2f}%\n".format(injury_by_sex.iloc[1,0]/injury_by_sex.iloc[2, 0]*100))

print("P(Gender = Female | Severity = Serious Injury) = {:.2f}%".format(injury_by_sex.iloc[0,1]/injury_by_sex.iloc[2,1]*100))
print("P(Gender = Male   | Severity = Serious Injury) = {:.2f}%\n".format(injury_by_sex.iloc[1,1]/injury_by_sex.iloc[2,1]*100))

print("P(Gender = Female | Severity = Slight Injury) = {:.2f}%".format(injury_by_sex.iloc[0,2]/injury_by_sex.iloc[2,2]*100))
print("P(Gender = Male   | Severity = Slight Injury) = {:.2f}%".format(injury_by_sex.iloc[1,2]/injury_by_sex.iloc[2,2]*100))

P(Gender = Female | Severity = Fatal) = 3.18%
P(Gender = Male   | Severity = Fatal) = 96.82%

P(Gender = Female | Severity = Serious Injury) = 6.03%
P(Gender = Male   | Severity = Serious Injury) = 93.97%

P(Gender = Female | Severity = Slight Injury) = 5.77%
P(Gender = Male   | Severity = Slight Injury) = 94.23%

Hypothesis 2: Is Speeding is the main Cause for Car Crashes?

accident_cause_by_injury = pd.crosstab(index=hypothesis_df['accident_cause'], columns=hypothesis_df['accident_severity'], margins=True)
accident_cause_by_injury #.sort_values(by = ["All"], ascending = False)

accident_severity	Fatal injury	Serious Injury	Slight Injury	All
accident_cause
Changing lane to the left	16	206	1251	1473
Changing lane to the right	23	260	1525	1808
Driving at high speed	2	31	141	174
Driving carelessly	22	209	1171	1402
Driving to the left	4	53	227	284
Driving under the influence of drugs	5	46	289	340
Drunk driving	0	3	24	27
Getting off the vehicle improperly	3	29	165	197
Improper parking	1	2	22	25
Moving Backward	26	162	949	1137
No distancing	20	303	1940	2263
No priority to pedestrian	5	95	621	721
No priority to vehicle	13	149	1045	1207
Other	7	64	385	456
Overloading	2	10	47	59
Overspeed	1	15	45	61
Overtaking	4	75	351	430
Overturning	2	23	124	149
Turnover	2	6	70	78
Unknown	0	2	23	25
All	158	1743	10415	12316

accidents_by_high_speed = (accident_cause_by_injury.loc["Overspeed", "All"] + 
                           accident_cause_by_injury.loc["Driving at high speed", "All"])

all_accidents = accident_cause_by_injury.loc["All", "All"]

print(accidents_by_high_speed)
print(all_accidents)

235
12316

print("Injuries(all types due to speeding): " + str(accidents_by_high_speed))
print("In percentage: {:.2f}%".format(accidents_by_high_speed/all_accidents*100))

Injuries(all types due to speeding): 235
In percentage: 1.91%

💡 Observations:

• Speeding is not the main factor for car accidents.

Hypothesis 3: Do weekends account for more accidents?

injury_by_day = pd.crosstab(index=hypothesis_df['day_of_week'], columns=hypothesis_df['accident_severity'], margins=True)
injury_by_day

accident_severity	Fatal injury	Serious Injury	Slight Injury	All
day_of_week
Friday	16	313	1712	2041
Monday	12	204	1465	1681
Saturday	37	245	1384	1666
Sunday	35	190	1242	1467
Thursday	22	272	1557	1851
Tuesday	17	257	1496	1770
Wednesday	19	262	1559	1840
All	158	1743	10415	12316

print('Accident percentage on weekdays:',round(sum([injury_by_day.iloc[i,3] for i in range(0,7) if injury_by_day.index[i] not in ['Saturday','Sunday']])/injury_by_day.iloc[7,3],2))
print('Accident percentage on weekends:',round(sum([injury_by_day.iloc[i,3] for i in range(0,7) if injury_by_day.index[i] in ['Saturday','Sunday']])/injury_by_day.iloc[7,3],2))

Accident percentage on weekdays: 0.75
Accident percentage on weekends: 0.25

💡 Observations:

• Its quite clear that there is no indication of more accidents happening on weekends

Hypothesis 4: Are there more dangerous injuries at night than in day?

injury_by_light_condition = pd.crosstab(index=hypothesis_df['light_condition'], columns=hypothesis_df['accident_severity'], margins=True)
injury_by_light_condition

accident_severity	Fatal injury	Serious Injury	Slight Injury	All
light_condition
Darkness - lights lit	66	465	2755	3286
Darkness - lights unlit	0	7	33	40
Darkness - no lighting	5	49	138	192
Daylight	87	1222	7489	8798
All	158	1743	10415	12316

sum(injury_by_light_condition.iloc[0:3, 0] + injury_by_light_condition.iloc[0:3, 1])

592

sol = 0
for i in [0,1]:
    for j in [0,1,2]:
        sol += injury_by_light_condition.iloc[j,i]

r = sol/(sum(injury_by_light_condition.iloc[i,3] for i in [0,1,2]))
print('Dangerous injuries at night:',round(r,2),'%')

r = sum(injury_by_light_condition.iloc[i,2] for i in [0,1,2])/sum(injury_by_light_condition.iloc[i,3] for i in [0,1,2])
print('Slight injuries at night:',round(r,2),'%')

Dangerous injuries at night: 0.17 %
Slight injuries at night: 0.83 %

💡 Observations:

• Although it does that in daylight there are more number of accidents but the dangerous injuries percentage are almost same either be it day or night.

Hypothesis 5: More accidents under some specific non-normal weather?

Without considering normal weather

injury_by_weather = pd.crosstab(index=hypothesis_df['weather_condition'], columns=hypothesis_df['accident_severity'], margins=True)
injury_by_weather.drop(['Normal','All'],axis=0)

accident_severity	Fatal injury	Serious Injury	Slight Injury	All
weather_condition
Cloudy	0	8	117	125
Fog or mist	0	1	9	10
Other	0	28	268	296
Raining	23	158	1150	1331
Raining and Windy	0	2	38	40
Snow	0	5	56	61
Unknown	0	51	241	292
Windy	0	16	82	98

💡 Observations:

• Without any analysis we can clearly see that rainy weather causes more accident and also all the fatal injuries have occured under non-normal conditions are there in rainy weather

⚙️ Preprocessing

df.isna().sum()

day_of_week                   0
driver_age                    0
driver_sex                    0
educational_level           741
vehicle_driver_relation     579
driving_experience          829
vehicle_type                950
vehicle_owner               482
service_year               3928
vehicle_defect             4427
accident_area               239
lanes                       385
road_allignment             142
junction_type               887
surface_type                172
road_surface_conditions       0
light_condition               0
weather_condition             0
collision_type              155
vehicles_involved             0
casualties                    0
vehicle_movement            308
casualty_class                0
casualty_sex                  0
casualty_age                  0
casualty_severity             0
casualty_work              3198
casualty_fitness           2635
pedestrian_movement           0
accident_cause                0
accident_severity             0
hour                          0
minute                        0
dtype: int64

def fill_na_with_distribution(df, cols):
    for col in cols:
        dist = df[col].value_counts(normalize=True)
        missing = df[col].isnull()
        if missing.sum() > 0:
            df.loc[missing, col] = np.random.choice(dist.index, size=missing.sum(), p=dist.values)
    return df

df = fill_na_with_distribution(df, ['casualty_class', 'casualty_sex', 'casualty_age', 'casualty_severity'])

# dropping columns that can cause imbalance while imputation
df.drop(columns = ['vehicle_defect', 'vehicle_driver_relation', 'casualty_work', 'casualty_fitness'], inplace=True)

impute_cols = [x for x in df.isna().sum()[df.isna().sum() != 0].index.tolist()]
for feat in impute_cols:
    mode = df[feat].mode()[0]
    df[feat].fillna(mode, inplace=True)

df.isna().sum() 

day_of_week                0
driver_age                 0
driver_sex                 0
educational_level          0
driving_experience         0
vehicle_type               0
vehicle_owner              0
service_year               0
accident_area              0
lanes                      0
road_allignment            0
junction_type              0
surface_type               0
road_surface_conditions    0
light_condition            0
weather_condition          0
collision_type             0
vehicles_involved          0
casualties                 0
vehicle_movement           0
casualty_class             0
casualty_sex               0
casualty_age               0
casualty_severity          0
pedestrian_movement        0
accident_cause             0
accident_severity          0
hour                       0
minute                     0
dtype: int64

⛓ Preprocessing

def ordinal_encoder(df, feats): 
    for feat in feats:    
        feat_val = list(np.arange(df[feat].nunique()))
        feat_key = list(df[feat].sort_values().unique())
        feat_dict = dict(zip(feat_key, feat_val))
        df[feat] = df[feat].map(feat_dict)
    return df

df = ordinal_encoder(df, df.drop(['accident_severity'], axis=1).columns)
df.shape

(12316, 29)

display_html(df.head(2), df.tail(2), df.sample(2))

	day_of_week	driver_age	driver_sex	educational_level	driving_experience	vehicle_type	vehicle_owner	service_year	accident_area	lanes	road_allignment	junction_type	surface_type	road_surface_conditions	light_condition	weather_condition	collision_type	vehicles_involved	casualties	vehicle_movement	casualty_class	casualty_sex	casualty_age	casualty_severity	pedestrian_movement	accident_cause	accident_severity	hour	minute
0	1	0	1	0	0	0	3	3	9	2	5	1	0	0	3	2	3	1	1	2	3	2	5	3	5	9	Slight Injury	17	1
1	1	1	1	4	3	11	3	2	6	4	5	1	0	0	3	2	8	1	1	2	3	2	5	3	5	16	Slight Injury	17	1

	day_of_week	driver_age	driver_sex	educational_level	driving_experience	vehicle_type	vehicle_owner	service_year	accident_area	lanes	road_allignment	junction_type	surface_type	road_surface_conditions	light_condition	weather_condition	collision_type	vehicles_involved	casualties	vehicle_movement	casualty_class	casualty_sex	casualty_age	casualty_severity	pedestrian_movement	accident_cause	accident_severity	hour	minute
12314	3	0	0	4	3	5	3	1	6	4	7	1	0	0	0	2	8	1	0	4	3	2	5	3	5	5	Slight Injury	13	11
12315	3	0	1	4	2	7	3	1	5	4	7	2	0	0	0	2	8	1	0	8	2	0	2	2	1	1	Slight Injury	13	11

	day_of_week	driver_age	driver_sex	educational_level	driving_experience	vehicle_type	vehicle_owner	service_year	accident_area	lanes	road_allignment	junction_type	surface_type	road_surface_conditions	light_condition	weather_condition	collision_type	vehicles_involved	casualties	vehicle_movement	casualty_class	casualty_sex	casualty_age	casualty_severity	pedestrian_movement	accident_cause	accident_severity	hour	minute
3969	4	4	0	1	2	0	0	1	7	6	5	1	0	0	3	7	8	1	0	4	0	1	1	1	5	10	Slight Injury	15	9
11228	3	1	1	6	1	5	3	5	7	6	5	7	0	0	3	2	8	1	2	2	1	0	3	2	5	12	Slight Injury	10	4

plt.figure(figsize=(22,17))
sns.set(font_scale=0.8)
sns.heatmap(df.corr(), annot=True, cmap=plt.cm.CMRmap_r)

<AxesSubplot: >

Upsampling

# Importing necessary libraries
from imblearn.over_sampling import SMOTE
from sklearn.model_selection import (
    cross_val_score,
    StratifiedKFold,
    KFold,
    train_test_split,
    GridSearchCV,
)
from sklearn.metrics import (
    accuracy_score,
    classification_report,
    recall_score,
    precision_score,
    f1_score,
    confusion_matrix,
)
from mlxtend.evaluate import mcnemar_table, mcnemar_tables
from mlxtend.plotting import checkerboard_plot, plot_decision_regions
from xgboost import XGBClassifier
from sklearn.ensemble import ExtraTreesClassifier, RandomForestClassifier
from sklearn.decomposition import PCA
from sklearn.tree import DecisionTreeClassifier
from sklearn.neighbors import KNeighborsClassifier
from sklearn.linear_model import LogisticRegression

X = df.drop('accident_severity', axis=1)
y = df['accident_severity']

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)
print(X_train.shape, X_test.shape, y_train.shape, y_test.shape)

(8621, 28) (3695, 28) (8621,) (3695,)

# upsampling using smote

counter = Counter(y_train)
counter

Counter({'Slight Injury': 7324, 'Serious Injury': 1191, 'Fatal injury': 106})

# upsampling using smote

counter = Counter(y_train)

print("=============================")

for k,v in counter.items():
    per = 100*v/len(y_train)
    print(f"Class= {k}, n={v} ({per:.2f}%)")

oversample = SMOTE()
X_train, y_train = oversample.fit_resample(X_train, y_train)

counter = Counter(y_train)

print("=============================")

for k,v in counter.items():
    per = 100*v/len(y_train)
    print(f"Class= {k}, n={v} ({per:.2f}%)")

print("=============================")

print("Upsampled data shape: ", X_train.shape, y_train.shape)

=============================
Class= Slight Injury, n=7324 (84.96%)
Class= Serious Injury, n=1191 (13.82%)
Class= Fatal injury, n=106 (1.23%)
=============================
Class= Slight Injury, n=7324 (33.33%)
Class= Serious Injury, n=7324 (33.33%)
Class= Fatal injury, n=7324 (33.33%)
=============================
Upsampled data shape:  (21972, 28) (21972,)

y_test = ordinal_encoder(pd.DataFrame(y_test, columns = ['accident_severity']), pd.DataFrame(y_test, columns = ['accident_severity']).columns)['accident_severity']
y_train = ordinal_encoder(pd.DataFrame(y_train, columns = ['accident_severity']), pd.DataFrame(y_train, columns = ['accident_severity']).columns)['accident_severity']

🧰 Baseline Modelling

def modelling(X_train, y_train, X_test, y_test, **kwargs):
    scores = {}
    models = []
    bvd = {}
    if 'xgb' in kwargs.keys() and kwargs['xgb']:
        xgb = XGBClassifier()
        xgb.fit(X_train._get_numeric_data(), np.ravel(y_train, order='C'))
        y_pred = xgb.predict(X_test._get_numeric_data())
        scores['xgb']= [accuracy_score(y_test, y_pred)]
        models.append(xgb)


    if 'rf' in kwargs.keys() and kwargs['rf']:
        rf = RandomForestClassifier(n_estimators=200)
        rf.fit(X_train, y_train)
        y_pred = rf.predict(X_test)
        scores['rf']= [accuracy_score(y_test, y_pred)]
        models.append(rf)

    if 'extree' in kwargs.keys() and kwargs['extree']:
        extree = ExtraTreesClassifier()
        extree.fit(X_train, y_train)
        y_pred = extree.predict(X_test)
        scores['extree'] = [accuracy_score(y_test, y_pred)]
        models.append(extree)
        
    return scores,models

scores,models = modelling(X_train,y_train, X_test, y_test, xgb=True, rf=True, extree=True)
print(scores)

{'xgb': [0.7948579161028417], 'rf': [0.7964817320703653], 'extree': [0.8056833558863329]}

y_pred_xgb = models[0].predict(X_test)
y_pred_rf = models[1].predict(X_test)
y_pred_ext = models[2].predict(X_test)


tb = mcnemar_tables(y_test, 
                    y_pred_xgb,
                    y_pred_rf, 
                    y_pred_ext)

brd = checkerboard_plot(tb['model_0 vs model_1'],
                        figsize=(5, 5),
                        fmt='%d',
                        col_labels=['model rf wrong', 'model rf right'],
                        row_labels=['model xgb wrong', 'model xgb right'])
plt.show()

brd = checkerboard_plot(tb['model_0 vs model_2'],
                        figsize=(5, 5),
                        fmt='%d',
                        col_labels=['model ext wrong', 'model ext right'],
                        row_labels=['model xgb wrong', 'model xgb right'])
plt.show()

brd = checkerboard_plot(tb['model_1 vs model_2'],
                        figsize=(5, 5),
                        fmt='%d',
                        col_labels=['model ext wrong', 'model ext right'],
                        row_labels=['model rf wrong', 'model rf right'])
plt.show()

def model_performance(model, y_test, y_hat) : 
    conf_matrix = confusion_matrix(y_test, y_hat)
    trace1 = go.Heatmap(z = conf_matrix  ,x = ["0 (pred)","1 (pred)", "2 (pred)"],
                        y = ["0 (true)","1 (true)", "2 (true)"],xgap = 2, ygap = 2, 
                        colorscale = 'Viridis', showscale  = False)

    #Show metrics
    Accuracy  =  accuracy_score(y_test, y_hat)
    Precision =  precision_score(y_test, y_pred, average= 'weighted')
    Recall    =  recall_score(y_test, y_pred, average= 'weighted')
    F1_score  =  f1_score(y_test, y_pred, average= 'weighted')

    show_metrics = pd.DataFrame(data=[[Accuracy , Precision, Recall, F1_score]])
    show_metrics = show_metrics.T

    colors = ['gold', 'lightgreen', 'lightcoral', 'lightskyblue']
    trace2 = go.Bar(x = (show_metrics[0].values), 
                   y = ['Accuracy', 'Precision', 'Recall', 'F1_score'], text = np.round_(show_metrics[0].values,4),
                    textposition = 'auto',
                   orientation = 'h', opacity = 0.8,marker=dict(
            color=colors,
            line=dict(color='#000000',width=1.5)))
 
     
    #plots
    model = model
    
    #Subplots
    fig = tls.make_subplots(rows=2, cols=1, print_grid=False, 
                          subplot_titles=('Confusion Matrix',
                                        'Metrics',
                                        ))
    
    fig.append_trace(trace1,1,1)
    fig.append_trace(trace2,2,1)
    
    fig['layout'].update(showlegend = False, title = '<b>Model performance report</b><br>'+str(model),
                        autosize = True, height = 800,width = 800,
                        plot_bgcolor = 'rgba(240,240,240, 0.95)',
                        paper_bgcolor = 'rgba(240,240,240, 0.95)',
                        # margin = dict(b = 100)
                        )
    fig.layout.titlefont.size = 14
    
    py.iplot(fig,filename='model-performance')

extree = ExtraTreesClassifier()
extree.fit(X_train, y_train)
y_pred = extree.predict(X_test)

extree.get_params()

{'bootstrap': False,
 'ccp_alpha': 0.0,
 'class_weight': None,
 'criterion': 'gini',
 'max_depth': None,
 'max_features': 'sqrt',
 'max_leaf_nodes': None,
 'max_samples': None,
 'min_impurity_decrease': 0.0,
 'min_samples_leaf': 1,
 'min_samples_split': 2,
 'min_weight_fraction_leaf': 0.0,
 'n_estimators': 100,
 'n_jobs': None,
 'oob_score': False,
 'random_state': None,
 'verbose': 0,
 'warm_start': False}

model_performance(extree,y_test, y_pred)

🛠️ Hyperparameter Tuning

gkf = KFold(n_splits=3, shuffle=True, random_state=42).split(X=X_train, y=y_train)

# A parameter grid for ETrees
params = {
    'n_estimators': range(100, 500, 100),  # [100,200,300,400,500]
    'ccp_alpha': [0.0, 0.1],
    'criterion': ['gini'],
    'max_depth': [5,11],
    'min_samples_split': [2,3],
}

extree_estimator = ExtraTreesClassifier()

gsearch = GridSearchCV(
    estimator= extree_estimator,
    param_grid= params,
    scoring='f1_weighted',
    n_jobs=-1,
    cv=gkf,
    verbose=1,
)


extree_model = gsearch.fit(X=X_train, y=y_train)
(gsearch.best_params_, gsearch.best_score_)

Fitting 3 folds for each of 32 candidates, totalling 96 fits

({'ccp_alpha': 0.0,
  'criterion': 'gini',
  'max_depth': 11,
  'min_samples_split': 2,
  'n_estimators': 300},
 0.8647603875807635)

gkf2 = KFold(n_splits=3, shuffle=True, random_state=101).split(X=X_train, y=y_train)

params2 = {
    'max_depth': [11,15],
    'min_samples_split': [2,3],
    'class_weight': ['balanced', None],

}

extree2 = ExtraTreesClassifier(ccp_alpha = 0.0,
                                criterion = 'gini',
                                max_depth = 11,
                                min_samples_split = 3)

gsearch2 = GridSearchCV(
    estimator= extree2,
    param_grid= params2,
    scoring='f1_weighted',
    n_jobs=-1,
    cv=gkf2,
    verbose=3,
)

extree_model2 = gsearch2.fit(X=X_train, y=y_train)

(gsearch2.best_params_, gsearch2.best_score_)

Fitting 3 folds for each of 8 candidates, totalling 24 fits

({'class_weight': None, 'max_depth': 15, 'min_samples_split': 2},
 0.9203765783132133)

extree_tuned = ExtraTreesClassifier(ccp_alpha = 0.0,
                                criterion = 'gini',
                                min_samples_split = 2,
                                class_weight = 'balanced',
                                max_depth = 15,
                                n_estimators = 400)

extree_tuned.fit(X_train, y_train)

ExtraTreesClassifier(class_weight='balanced', max_depth=15, n_estimators=400)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

y_pred_tuned = extree_tuned.predict(X_test)

print(extree_tuned.feature_importances_)

[0.0565462  0.05789308 0.01420086 0.02952284 0.04402301 0.03613448
 0.0199071  0.03517829 0.02860482 0.03771439 0.01405304 0.0453975
 0.01112801 0.04420206 0.06420246 0.0178615  0.03507673 0.0687727
 0.05905148 0.02040236 0.02924959 0.03142726 0.02946688 0.02458844
 0.01077821 0.04141331 0.04219812 0.05100529]

feat_importances = pd.Series(extree_tuned.feature_importances_, index=X.columns)
plt.figure(figsize=(12,12))
myexplode = [0.12,0,0,0,0,0,0,0,0,0]
plt.pie(feat_importances.nlargest(10),labels=feat_importances.nlargest(10).index, autopct='%.0f%%',explode= myexplode,
        textprops={'fontsize': 16})
centre_circle = plt.Circle((0,0),0.10,fc='white')
fig = plt.gcf()
fig.gca().add_artist(centre_circle)
plt.title('Important features considered by Extra Trees Classifier',fontsize=20)
plt.show()

Explainable AI

shap.initjs()

X_sample = X_train.sample(100)
X_sample

	day_of_week	driver_age	driver_sex	educational_level	driving_experience	vehicle_type	vehicle_owner	service_year	accident_area	lanes	road_allignment	junction_type	surface_type	road_surface_conditions	light_condition	weather_condition	collision_type	vehicles_involved	casualties	vehicle_movement	casualty_class	casualty_sex	casualty_age	casualty_severity	pedestrian_movement	accident_cause	hour	minute
10697	3	1	1	4	0	0	3	0	8	5	3	7	0	1	0	2	8	1	1	2	1	0	0	2	5	3	18	5
303	0	1	1	4	4	0	3	3	9	6	5	1	0	0	3	2	1	1	0	2	0	1	3	2	5	12	9	9
5495	0	0	1	4	2	4	3	5	6	4	5	7	0	0	3	2	8	2	0	0	3	2	5	3	5	1	17	4
10431	2	0	1	3	1	6	2	5	4	4	5	1	0	0	0	2	8	1	2	3	0	0	0	2	5	2	20	2
18118	3	0	1	4	3	0	2	4	6	0	5	1	3	0	3	2	8	0	0	2	3	2	5	3	5	12	15	10
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
4105	5	0	1	0	2	0	3	2	6	2	5	7	0	0	3	2	8	1	1	2	3	2	5	3	5	16	7	10
15682	5	2	1	2	1	8	3	3	5	2	5	6	0	0	2	2	8	0	0	2	3	2	5	3	5	9	10	8
14573	3	0	1	4	1	0	3	5	6	4	5	1	0	2	3	3	0	1	2	4	0	1	0	2	5	4	15	4
4813	1	0	1	4	1	15	3	0	9	2	5	7	0	0	3	2	8	2	0	2	1	1	0	2	5	0	11	10
7394	2	4	1	4	3	5	3	2	7	4	5	1	0	3	0	4	8	2	0	2	3	2	5	3	5	10	22	0

100 rows × 28 columns

shap_values = shap.TreeExplainer(extree_tuned).shap_values(X_sample)

shap.summary_plot(shap_values, X_sample, plot_type="bar")

shap.force_plot(shap.TreeExplainer(extree_tuned).expected_value[0],
                shap_values[0][:], 
                X_sample)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

print(y_pred_tuned[50])
shap.force_plot(shap.TreeExplainer(extree_tuned).expected_value[0], shap_values[1][50], X_sample.iloc[50])

2

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

i=13
print(y_pred_tuned[i])
shap.force_plot(shap.TreeExplainer(extree_tuned).expected_value[0], shap_values[0][i], X_sample.values[i], feature_names = X_sample.columns)

1

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

print(y_pred_tuned[10])
row = 10
shap.waterfall_plot(shap.Explanation(values=shap_values[0][row], 
                                              base_values=shap.TreeExplainer(extree_tuned).expected_value[0], data=X_sample.iloc[row],  
                                         feature_names=X_sample.columns.tolist()))

2

shap.dependence_plot('day_of_week', shap_values[2], X_sample)

shap.dependence_plot('driver_age', shap_values[2], X_sample)

print(y_pred_tuned[10])
shap.decision_plot(shap.TreeExplainer(extree_tuned).expected_value[0], 
                   shap_values[2][:10], 
                   feature_names=X_sample.columns.tolist())

2

# upsampling using smote
y = df["accident_severity"]
X = df.drop(["accident_severity"], axis=1)

counter = Counter(y)

for k, v in counter.items():
    dist = v / len(y) * 100
    print(f"Class={k}, n={v} ({per:.2f}%)")

models = []
models.append(('KNN', KNeighborsClassifier()))
models.append(('LOG', LogisticRegression()))
models.append(('DTC', DecisionTreeClassifier()))
models.append(('RFC', RandomForestClassifier()))

names = []
results = []
for name, model in models:
    fold = KFold(n_splits=10)
    score = cross_val_score(model, X, y, cv=fold, scoring='accuracy')
    names.append(name)
    results.append(score)

plotdict = dict(zip(names, results))
for k, v in plotdict.items():
    print(f"{k}, {round(v.mean(), 5)}")

top10 = ['hour', "casualties", "day_of_week", "accident_cause", "vehicles_involved", "vehicle_type", 
         "driver_age", "accident_area", "driving_experience", "lanes"]

df10 = df[top10]
df10

X_train_new, X_test_new, y_train_new, y_test_new = train_test_split(df10, y, test_size=0.3, random_state=42)
print(X_train_new.shape, X_test_new.shape, y_train_new.shape, y_test_new.shape)

model = RandomForestClassifier()

model.fit(X_train_new, y_train_new)

y_pred_new = model.predict(X_test_new)
score_new = accuracy_score(y_test_new, y_pred_new)
print("Accuracy score: ", score_new)

joblib.dump(model, 'random_forest_final.joblib')

df.shape

(12316, 29)

def imbalanced_features(df):
    mod_imb=[] # moderately imbalanced
    hi_imb=[] # highly imbalanced

    for col in df.columns:
        try:
            if (df[col].value_counts()/df.shape[0]).max() > 0.9:
                hi_imb.append(col)
            elif (df[col].value_counts()/df.shape[0]).max() > 0.7 and col not in hi_imb:
                mod_imb.append(col)
        except Exception as e:
            print(f"Couldn't check \033[1m{col}\033[0m. ", e, "\n")

    print("="*20, "\033[1m Imbalanced features\033[0m", "="*20)
    print("No of moderately imbalanced features (75%+ data on a single class): ", len(mod_imb), "\n")
    print(mod_imb, "\n")
    print("No of highly imbalanced features (90%+ data on a single class): ", len(hi_imb), "\n")
    print(hi_imb)
imbalanced_features(df)

====================  Imbalanced features ====================
No of moderately imbalanced features (75%+ data on a single class):  7 

['vehicle_owner', 'road_allignment', 'road_surface_conditions', 'light_condition', 'weather_condition', 'collision_type', 'accident_severity'] 

No of highly imbalanced features (90%+ data on a single class):  3 

['driver_sex', 'surface_type', 'pedestrian_movement']

# upsampling using smote
y = df['accident_severity']
X = df.drop('accident_severity', axis=1)

counter = Counter(y)

for k,v in counter.items():
    per = 100*v/len(y)
    print(f"Class= {k}, n={v} ({per:.2f}%)")

Class= Slight Injury, n=10415 (84.56%)
Class= Serious Injury, n=1743 (14.15%)
Class= Fatal injury, n=158 (1.28%)

oversample = SMOTE()
X, y = oversample.fit_resample(X, y)

counter = Counter(y)

for k,v in counter.items():
    per = 100*v/len(y)
    print(f"Class= {k}, n={v} ({per:.2f}%)")

Class= Slight Injury, n=10415 (33.33%)
Class= Serious Injury, n=10415 (33.33%)
Class= Fatal injury, n=10415 (33.33%)

df = pd.concat([X, y], axis=1)
print("Upsampled data shape: ", df.shape)

Upsampled data shape:  (31245, 29)

# selecting a good baseline model using cross validation

models = []
models.append(('KNN', KNeighborsClassifier()))
models.append(('LOG', LogisticRegression()))
models.append(('DTC', DecisionTreeClassifier()))
models.append(('RFC', RandomForestClassifier()))

names = []
results = []
for name, model in models:
    fold = KFold(n_splits=10)
    score = cross_val_score(model, X, y, cv=fold, scoring='accuracy')
    names.append(name)
    results.append(score)

plotdict = dict(zip(names, results))

for k,v in plotdict.items():
    print(f"{k}: {round(v.mean(),5)}")

KNN: 0.79581
LOG: 0.45408
DTC: 0.81608
RFC: 0.91164

model = RandomForestClassifier()

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)
print(X_train.shape, X_test.shape, y_train.shape, y_test.shape)

(21871, 28) (9374, 28) (21871,) (9374,)

# fitting model
model.fit(X_train, y_train)

RandomForestClassifier()

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

# predicting
y_pred = model.predict(X_test)

score = accuracy_score(y_test, y_pred)
print("Accuracy: ", score)

Accuracy:  0.9270322167697888

# getting feature importance
model.feature_importances_

array([0.05976835, 0.04902253, 0.00852862, 0.02963963, 0.04457838,
       0.05237394, 0.01392963, 0.03375527, 0.04351982, 0.04510356,
       0.01514861, 0.04193056, 0.00787389, 0.0302327 , 0.0413342 ,
       0.0184541 , 0.03003746, 0.05763699, 0.06237717, 0.03179664,
       0.0196426 , 0.01589697, 0.02627878, 0.01208849, 0.01006779,
       0.05598957, 0.07489044, 0.06810332])

df_importance = pd.DataFrame()
df_importance['Features'] = X.columns
df_importance['Importance'] = model.feature_importances_
plt.figure(figsize=(10, 12))
sns.barplot(data = df_importance.sort_values("Importance", ascending=False), y='Features', x='Importance');

# selecting top 10 features
top10 = list(df_importance.sort_values("Importance", ascending=False)['Features'].head(10).values)
top10

['hour',
 'minute',
 'casualties',
 'day_of_week',
 'vehicles_involved',
 'accident_cause',
 'vehicle_type',
 'driver_age',
 'lanes',
 'driving_experience']

df_top10 = df[['lanes',
 'minute',
 'casualties',
 'day_of_week',
 'vehicles_involved',
 'accident_cause',
 'vehicle_type',
 'driver_age',
 'lanes',
 'driving_experience']]
df_top10

	lanes	minute	casualties	day_of_week	vehicles_involved	accident_cause	vehicle_type	driver_age	lanes	driving_experience
0	2	1	1	1	1	9	0	0	2	0
1	4	1	1	1	1	16	11	1	4	3
2	6	1	1	1	1	0	5	0	6	0
3	6	2	1	3	1	1	11	0	6	2
4	6	2	1	3	1	16	0	0	6	1
...	...	...	...	...	...	...	...	...	...	...
31240	2	6	0	2	1	0	7	0	2	3
31241	2	1	0	3	0	11	5	1	2	1
31242	4	9	2	4	1	9	0	2	4	0
31243	3	10	0	3	1	10	0	0	3	1
31244	2	3	0	1	0	2	14	3	2	0

31245 rows × 10 columns

X_train_new, X_test_new, y_train_new, y_test_new = train_test_split(df_top10, y, test_size=0.3, random_state=42)
print(X_train_new.shape, X_test_new.shape, y_train_new.shape, y_test_new.shape)

(21871, 10) (9374, 10) (21871,) (9374,)

# fitting model
model.fit(X_train_new, y_train_new)

RandomForestClassifier()

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

# predicting
y_pred_new = model.predict(X_test_new)
score_new = accuracy_score(y_test_new, y_pred_new)
print("Accuracy: ", score_new)

Accuracy:  0.8595050138681459

joblib.dump(model, 'random_forest_final.joblib')

['random_forest_final.joblib']

💡 Observations:

A

💡 Observations:

A

💡 Observations:

A

💡 Observations:

A